AI reasoning is moving into the era of heterogeneous computing power! Microsoft Maia 300 accelerates the rise of self-developed silicon, and TSM.US (TSM.US) and MRVL.US (MRVL.US) welcome the ASIC growth cycle

The Zhitong Finance App learned that, citing news reports from people familiar with the matter, that when Microsoft (MSFT.US), one of the US tech giants, plans to launch its latest self-developed AI accelerator Maia 300 soon, Wedbush Securities, a well-known Wall Street investment institution, believes that this latest market trend may be very favorable to focus on AI ASIC and data center optical interconnect chips (MRVL.US) and TSM.US (TSM.US). Microsoft will probably announce the next-generation self-developed AI ASIC chip MAIA 300 in September 2026. It is currently in negotiations with TSMC to lock in production capacity for more than 300,000 chips in 2027, and hopes to obtain production capacity of more than 1 million chips in the long term.

Microsoft made it clear in a statement that its MAIA self-developed AI ASIC deployment ultimately faces huge AI inference workload requirements measured in gigawatts (GW). To a large extent, this means that hyperscale cloud computing vendors are completely upgrading from “group procurement of GPUs” to heterogeneous AI computing power infrastructure where Nvidia AI GPU+AMD AI GPU+self-developed ASIC/XPU coexist: AI training operator processes and cutting-edge AI workloads with the most complex architectures and the fastest changing rate still require AI GPU clusters, and large-scale AI inference workloads around mature and open source AI models, Copilots, and AI agents (AI agents) are becoming more and more suitable A dedicated self-developed AI chip.

Obviously, Microsoft is trying to follow the Google TPU commercial computing power leasing and computing power sales route of Google, another cloud computing giant, rather than simply using it for Microsoft's internal AI training or inference process. But currently it's more like renting Azure cloud computing power, rather than directly selling raw chips.” Microsoft has made it clear that the Maia 200 is increasing the AI inference computing capacity of the Azure cloud computing platform that customers can actually put into use, while opening Maia SDK previews to developers, AI startups, and academic institutions.

Overall, Microsoft's last-generation self-developed AI chip, MaiA 200, has entered the Azure cloud computing production environment from an in-house development chip and has begun to transform into AI inference computing power that customers can actually consume; however, it should not be directly interpreted at this stage as that customers can already rent an instance of the MaiA 200 chip on a large scale and separately like leasing an Nvidia Blackwell, H100, or Google TPU.

More importantly, Reuters recently reported that Microsoft is trying to persuade large cloud customers, including Anthropic, to adopt the Maia 300, which shows that Maia is no longer just a Microsoft 365 CoPilot or a closed self-developed chip for internal OpenAI workloads.

Microsoft Maia 300 impacted AI inference costs, and TSMC and Maywell are the biggest beneficiary chain?

Matt Bryson, senior analyst at Wedbush, wrote in a report to customers: “Until now, Microsoft's Maia self-developed AI chip project has failed to meet expectations, and overall, Microsoft's execution in the hardware field has been lacking in the past.”

“Having said that, Microsoft's AI chip self-development work is nearing an important point in the iteration of multiple generations of products, and judging from historical experience, the positive impact of chip development often starts to make real progress at this stage — for example, it takes 4 to 5 years or more, and progresses to third-generation products. If large-scale AI computing power infrastructure orders are eventually formed, the chip IP and chip design partner Maywell of Microsoft's self-developed AI chip and TSMC, which OEM the chip, should be direct beneficiaries of the success of Microsoft's self-developed AI chip.”

Analyst Bryson also pointed out that the Maia 300 may also have a certain impact on Nvidia (NVDA.US)'s stock price and basic outlook, because it will become another customized self-developed AI chip project by cloud giants that aims to reduce the cost of enterprise-side AI inference after Amazon and Google developed their own AI chip systems. However, Bryson added that Nvidia may not be greatly affected because the company has been “excellent” in “ensuring sufficient supply scale” and the general AI computing power base.

According to media reports, the Maia 300 could be released as early as fall this year. Microsoft has negotiated with TSMC to lock in manufacturing capacity for more than 300,000 of these chips, which is scheduled to be delivered in 2027. Maia 200 was released in January of this year, and the early version of the Maia self-developed AI chip project was first launched as early as November 2023.

Microsoft and Mywell did not immediately respond to media requests for comment.

Microsoft wants to copy Google's TPU path! Maia 300 points to the Azure computing power base, and TSMC and customized chip chains ushered in a new growth cycle

Looking at the underlying architecture, it is not “absolute computing power” that Microsoft MAIA actually has an advantage over Nvidia GPUs, but rather unit economy under extremely large, repetitive inference loads — cost per token, token throughput per watt (Tokens/Watt), and total cost of ownership (TCO) of AI servers.

GPUs must balance training, inference, scientific computing, and a large number of dynamic workloads, so the biggest moat is generalization+CUDA/TensorRT software ecosystem+NVLink-level cluster expansion capability; while MAIA is a customized self-developed AI accelerator based on Microsoft's software/hardware co-design benchmark system around Azure workloads, which can concentrate more transistors, power budget, and on-chip area into the low-precision tensor computation, memory capacity/bandwidth, and data handling that Transformer inference really requires.

The Maia 200 already uses TSMC's 3 nm manufacturing process (TSMC 3nm), native FP4/FP8 Tensor Core, 216GB HBM3e, 7Tb/s bandwidth, and 272MB on-chip SRAM. Microsoft has specially strengthened DMA, NoC, and two-level scale-up networks, essentially optimizing the architecture for the increasingly serious memory wall (Memory Wall), KV Cache, and data movement costs in the inference phase; Microsoft's latest financial report says the Maia 200 is compared to its fleet The latest generation of hardware has achieved a 30% increase in performance per dollar, and its self-developed MAI model has improved performance per watt by about 40% when running on the MAIA 200. This is the most dangerous competitiveness of ASIC/XPU in the age of reasoning: when billions of similar token generation tasks can be highly standardized, the flexibility premium of Nvidia-led “universal GPUs” may not necessarily be worth paying for every inference request.

If the Maia 300 is implemented at the current pace and scale of media reports, it can be called a “major structural benefit” for the global AI semiconductor investment theme, but it also means further expansion of AI computing power infrastructure TAM (overall market size) and re-stratification of profit pools — that is, TSMC's advanced process production capacity, advanced process industry chain, customized ASIC design, HBM/NAND/data center level DRAM, advanced packaging, high-speed optical interconnection and data center networks. However, for the AI GPU technology route dominated by Nvidia and AMD Market expansion expectations are likely to be negative.

Chris Caso, a strategist from Wolfe Research, a top Wall Street investment agency, pointed out that the Philadelphia Semiconductor Index (SOXX) doubled in the previous three months and then fell about 25% from its high. The recent weakness is more like a reset of expectations after a sharp rise. Caso anticipates that demand for artificial intelligence chips will still exceed supply until at least 2028. He believes that the market's previously feared slowdown in capital expenditure for hyperscale cloud computing has not actually occurred, and that the competitive trend surrounding AI agents has instead left hyperscale cloud vendors “no choice not to invest.”

According to a recent research report led by Brian Nowak, a senior analyst at Morgan Stanley, the 2027/2028 capital expenditure forecasts for the five largest hyperscale cloud computing and vendors (Meta, Amazon, Microsoft, Google, SpaceX) in the global market (Meta, Amazon, Microsoft, Google, SpaceX) have been significantly raised again, reaching approximately $1.2 trillion and $1.4 trillion, respectively. The agency's capital expenditure forecast for major US tech giants in 2026 was drastically raised from 433 billion US dollars a year ago to 805 billion US dollars.

But that doesn't mean Maia can replace Nvidia in the short term. On the contrary, inference competition has been upgraded from “chip peak FLOPS” to “complete AI factory efficiency,” and Nvidia's strongest barriers are still CUDA, TensorRT-LLM, Dynamo, mature kernel libraries, and the integrated ecosystem of GPU-NVLink-Network-Software. TensorRT-LLM has continuously optimized inference critical paths such as FP8/FP4 quantization, Paged KV Cache, Speculative Decoding, and Expert Parallelism. Therefore, if Microsoft wants to actually turn Maia into a large-scale third-party platform, the biggest challenge is not designing a chip, but making the compiler, kernel, scheduler, model compatibility, and developer ecosystem mature enough.

According to institutions such as Morgan Stanley and Wedbush Securities that continue to be optimistic about the investment prospects of the AI computing power industry chain, the almost limitless cutting-edge computing power in the AI reasoning era and the computing power demand surrounding AI agents enabled AI ASICs to grow into an important component of the second trillion-level computing power ecosystem without destroying GPU demand — instead further strengthening the AI computing power industry chain investment logic that “the AI semiconductor supercycle is not a single GPU cycle, but an entire data center silicon content increase cycle”.

In terms of TSMC's stock price outlook, the “king of chip foundry,” Barclays Bank analyst Simon Coles recently raised TSMC's target price from 625 US dollars to 650 US dollars and maintain “additional holdings”; if this price is achieved, it means that TSMC's market value may jump from about 2.17 trillion US dollars to about 3.37 trillion US dollars.

Coles' core bullish logic can be summed up as follows: Global AI is manufacturing an “advanced logic wafer supply battle” — its Asian supply chain research has observed strong competition among customers over advanced logic wafer production capacity, while TSMC has also grasped the two major AI computing power bottlenecks of N2/N3 advanced manufacturing processes and CowOS advanced packaging, and this scarcity is being transformed into strong quantitative pricing power. In other words, the $650 bull market scenario is not only about “Nvidia continues to sell GPUs,” but that all heterogeneous AI computing power clusters such as Nvidia/AMD GPU+Google TPU+AWS Trainium+Microsoft Maia+AMD GPU will eventually require advanced manufacturing processes and advanced packaging. As a result, TSMC is becoming the semiconductor asset closest to the “computing power infrastructure transit toll gate” in the AI reasoning era.

In terms of Maywell's stock price outlook, the most aggressive bullish point is KeyBANC analyst John Vinh's target price of $400, which was raised again from $385 on July 14 and maintained; compared to the current trading price of $208.56, this means there is room for growth of about 91.8%. Based on the current estimate of about 875.8 million shares in circulation, $400 means Maywell's potential market value is about US$35.3 billion. Vinh's core bullish logic closely revolves around the “cloud giant's heterogeneous computing power+AI reasoning era” — that is, Marvell (Marvell) is simultaneously receiving two growth curves of AI ASIC/XPU (Custom Silicon) and high-speed interconnect (Optical/Networking) customized by hyperscale cloud vendors.

The $400 raised by KeyBanc analyst Vinh is not simply betting on the MAIA 300 chip, but is betting on an era where cloud giants such as Microsoft, Google, and Amazon switch from “purchasing Nvidia GPUs” to heterogeneous computing power of GPU+ self-developed ASICs, Marvell will further upgrade from an AI interconnection supplier to a core “arms dealer” for cloud giants to develop their own silicon.